Dispositivo para identificação de atividades físicas baseado em sensores inerciais e redes neurais

Abstract

The main objective of this work is to develop a prototype that uses inertial sensors and neural networks to identify physical activities, with and without dumbbells, which can be used in physical evaluation centers and hospital environments. The device consists of two inertial sensors, which are placed on the wrist and arm of the volunteers and a central one that captures and transmits data from the sensors to a computer for storage. Data were collected from 7 volunteers, with an average age of 27.4 ± 7.1 years, performing 15 physical activities. The collected accelerometer and gyroscope signal data are processed and displayed as inputs to a Multilayer Neural Network (Multilayer Perceptron) with 1 hidden layer. For these stages, several configurations were evaluated for the processing blocks and the neural network of the Edge Impulse platform, in order to seek maximum accuracy during the testing stage. Accuracy values of 98.91% were obtained with data from both sensors, 94.77% with the sensor on the arm and 97.77% due to signals from the sensor placed on the wrist. For these results, the Neural Network had 480 neurons in its only hidden layer.